More Flexible Monad Transformer OR Bad Coding Style

aditya siram

9 Aug 2010 9 Aug '10

7:05 p.m.

Hi all, I was experimenting with monad transformers and realized that the stacking order of the monads can remain unknown until it is used. Take for example the following code: import "mtl" Control.Monad.State import "mtl" Control.Monad.Writer import "mtl" Control.Monad.Identity test :: (MonadWriter [Char] m, Num t, MonadState t m) => m () test = do put 1 tell "hello" main = do x <- return $ runIdentity $ runStateT (runWriterT test) 1 -- test :: WriterT String (StateT Int Identity) y <- return $ runIdentity $ runWriterT $ runStateT test 1 -- test :: StateT Int (WriterT String Identity) z <- runWriterT $ runStateT test 1 -- test :: StateT Int (WriterT String IO) (((), Int), String) print x print y print z *Main> main (((),"hello"),1) (((),1),"hello") (((),1),"hello") Until test is called in 'main' we don't know the order of monads. In fact even the base monad is not know. All we know is that it uses the State and Writer monad. In each call to 'test' in main we can determine the stacking order and the base monad yielding different results. This seems to be a more flexible way of using monad transformers but I haven't seen this in code before so is there anything wrong with this style? -deech

Attachments:

attachment.html (text/html — 2.9 KB)

Show replies by date

Edward Z. Yang

9 Aug 9 Aug

7:37 p.m.

Excerpts from aditya siram's message of Mon Aug 09 15:05:14 -0400 2010:

...

Until test is called in 'main' we don't know the order of monads. In fact even the base monad is not know. All we know is that it uses the State and Writer monad. In each call to 'test' in main we can determine the stacking order and the base monad yielding different results. This seems to be a more flexible way of using monad transformers but I haven't seen this in code before so is there anything wrong with this style?

There are two points here: * You have observed that running the monads in different orders results in different values: however, in the case of the examples you've chosen, the essential character of the result is still the same. This is not universally true: only some monads "commute", so to speak. So you definitely do not want to leave the ordering implicit if order matters. But when the monads are commutative... * The general feeling on this issue, as I've heard from Don, is that generic types for specifying monad stacks are too fragile: due to the extra generality, it's a lot easier to write code that typechecks but doesn't do what you actually want it to. A decent compromise might be to use a concrete stack, and offer a utility function that "lifts" it into the generic monad typeclasses. Cheers, Edward

Gábor Lehel

7:39 p.m.

New subject: More Flexible Monad Transformer OR Bad Coding Style

Actually, while I haven't even used monad transformers before (just read about them a lot), I was thinking that something like this might be the way to solve the lift . lift . lift . lift . foo problem on the one hand, and by wrapping the 'contents' (e.g. the environment of a reader monad) of every level of the stack in a unique newtype (if the type isn't otherwise unique), the problem of "what if I want to use the same transformer more than once, how do I disambiguate them". (Do I have roughly the right idea?) On Mon, Aug 9, 2010 at 9:05 PM, aditya siram wrote:

...

Hi all, I was experimenting with monad transformers and realized that the stacking order of the monads can remain unknown until it is used. Take for example the following code:

import "mtl" Control.Monad.State import "mtl" Control.Monad.Writer import "mtl" Control.Monad.Identity

test :: (MonadWriter [Char] m, Num t, MonadState t m) => m () test = do put 1 tell "hello"

main = do x <- return $ runIdentity $ runStateT (runWriterT test) 1 -- test :: WriterT String (StateT Int Identity) y <- return $ runIdentity $ runWriterT $ runStateT test 1 -- test :: StateT Int (WriterT String Identity) z <- runWriterT $ runStateT test 1 -- test :: StateT Int (WriterT String IO) (((), Int), String) print x print y print z

*Main> main (((),"hello"),1) (((),1),"hello") (((),1),"hello")

Until test is called in 'main' we don't know the order of monads. In fact even the base monad is not know. All we know is that it uses the State and Writer monad. In each call to 'test' in main we can determine the stacking order and the base monad yielding different results. This seems to be a more flexible way of using monad transformers but I haven't seen this in code before so is there anything wrong with this style?

-deech

_______________________________________________ Haskell-Cafe mailing list Haskell-Cafe@haskell.org http://www.haskell.org/mailman/listinfo/haskell-cafe

-- Work is punishment for failing to procrastinate effectively.

Edward Z. Yang

7:47 p.m.

Excerpts from Gábor Lehel's message of Mon Aug 09 15:39:49 -0400 2010:

...

Actually, while I haven't even used monad transformers before (just read about them a lot), I was thinking that something like this might be the way to solve the lift . lift . lift . lift . foo problem on the one hand, and by wrapping the 'contents' (e.g. the environment of a reader monad) of every level of the stack in a unique newtype (if the type isn't otherwise unique), the problem of "what if I want to use the same transformer more than once, how do I disambiguate them". (Do I have roughly the right idea?)

In fact, what you describe is already in use by current monad transformer libraries. :-) Take for example StateT in mtl [1]: MonadWriter w m => MonadWriter w (StateT s m) MonadError e m => MonadError e (StateT s m) Monad m => MonadState s (StateT s m) MonadReader r m => MonadReader r (StateT s m) MonadTrans (StateT s) Monad m => Monad (StateT s m) Monad m => Functor (StateT s m) MonadFix m => MonadFix (StateT s m) MonadPlus m => MonadPlus (StateT s m) MonadIO m => MonadIO (StateT s m) MonadCont m => MonadCont (StateT s m) By default, it comes with all of these instances, so that if you use a function that relies on say MonadReader and not Reader, no lifting is necessary. However, newtyping every reader monad because you want to use multiple copies of them isn't as good, because you still have to manually add all of these instances. Cheers, Edward [1] http://hackage.haskell.org/packages/archive/mtl/1.1.0.2/doc/html/Control-Mon...

Gábor Lehel

7:59 p.m.

New subject: More Flexible Monad Transformer OR Bad Coding Style

2010/8/9 Edward Z. Yang :

...

Excerpts from Gábor Lehel's message of Mon Aug 09 15:39:49 -0400 2010:

...
Actually, while I haven't even used monad transformers before (just read about them a lot), I was thinking that something like this might be the way to solve the lift . lift . lift . lift . foo problem on the one hand, and by wrapping the 'contents' (e.g. the environment of a reader monad) of every level of the stack in a unique newtype (if the type isn't otherwise unique), the problem of "what if I want to use the same transformer more than once, how do I disambiguate them". (Do I have roughly the right idea?)

In fact, what you describe is already in use by current monad transformer libraries. :-) Take for example StateT in mtl [1]:

MonadWriter w m => MonadWriter w (StateT s m) MonadError e m => MonadError e (StateT s m) Monad m => MonadState s (StateT s m) MonadReader r m => MonadReader r (StateT s m) MonadTrans (StateT s) Monad m => Monad (StateT s m) Monad m => Functor (StateT s m) MonadFix m => MonadFix (StateT s m) MonadPlus m => MonadPlus (StateT s m) MonadIO m => MonadIO (StateT s m) MonadCont m => MonadCont (StateT s m)

By default, it comes with all of these instances, so that if you use a function that relies on say MonadReader and not Reader, no lifting is necessary. However, newtyping every reader monad because you want to use multiple copies of them isn't as good, because you still have to manually add all of these instances.

What I meant was to make certain that the 'r' in `ReaderT r m a`, say, is a unique type in every case, so that when you write `foo :: MonadReader SomethingOrOther m => m a -> m a` (for example), it's always unambiguous which MonadReader in the stack you might've meant. (If it's not... would you get an overlapping instances error, or what?)

...

Cheers, Edward

[1] http://hackage.haskell.org/packages/archive/mtl/1.1.0.2/doc/html/Control-Mon...

-- Work is punishment for failing to procrastinate effectively.

Gregory Crosswhite

8:37 p.m.

I've never used this myself, but the package mtlx seems to offer one possible solution to this problem by tagging the monad transformers with index types: http://hackage.haskell.org/package/mtlx Cheers, Greg On 08/09/10 12:39, Gábor Lehel wrote:

...

Actually, while I haven't even used monad transformers before (just read about them a lot), I was thinking that something like this might be the way to solve the lift . lift . lift . lift . foo problem on the one hand, and by wrapping the 'contents' (e.g. the environment of a reader monad) of every level of the stack in a unique newtype (if the type isn't otherwise unique), the problem of "what if I want to use the same transformer more than once, how do I disambiguate them". (Do I have roughly the right idea?)

On Mon, Aug 9, 2010 at 9:05 PM, aditya siram wrote:

...
Hi all, I was experimenting with monad transformers and realized that the stacking order of the monads can remain unknown until it is used. Take for example the following code:

import "mtl" Control.Monad.State import "mtl" Control.Monad.Writer import "mtl" Control.Monad.Identity

test :: (MonadWriter [Char] m, Num t, MonadState t m) => m () test = do put 1 tell "hello"

main = do x <- return $ runIdentity $ runStateT (runWriterT test) 1 -- test :: WriterT String (StateT Int Identity) y <- return $ runIdentity $ runWriterT $ runStateT test 1 -- test :: StateT Int (WriterT String Identity) z <- runWriterT $ runStateT test 1 -- test :: StateT Int (WriterT String IO) (((), Int), String) print x print y print z

*Main> main (((),"hello"),1) (((),1),"hello") (((),1),"hello")

Until test is called in 'main' we don't know the order of monads. In fact even the base monad is not know. All we know is that it uses the State and Writer monad. In each call to 'test' in main we can determine the stacking order and the base monad yielding different results. This seems to be a more flexible way of using monad transformers but I haven't seen this in code before so is there anything wrong with this style?

-deech

_______________________________________________ Haskell-Cafe mailing list Haskell-Cafe@haskell.org http://www.haskell.org/mailman/listinfo/haskell-cafe

Job Vranish

7:42 p.m.

For monads like StateT, WriterT, ReaderT, the order doesn't matter (except perhaps for some pesky performance details). However, for monad transformers like ErrorT or ListT, the order _does_ matter. The code you have there is perfectly fine, sometimes the added generality can be quite handy (especially if you have your own MonadState'esk type classes). The two major drawbacks to this approach (that I can think of off the top of my head) are: 1) Rather large and complicated contexts on quite a few of your functions 2) Can lead to nearly indecypherable error messages Personally, I try to avoid multiparameter typeclasses whenever possible; I've found them to be more trouble than they are worth. My advice would be to leave the code general if the code actually does something general (it actually has more than one use case) and give the code a fixed signature if the code really one has just one purpose (even if ghci can infer a general type for you). This is just a personal preference, but it seems to work well for me :) - Job On Mon, Aug 9, 2010 at 3:05 PM, aditya siram wrote:

...

Hi all, I was experimenting with monad transformers and realized that the stacking order of the monads can remain unknown until it is used. Take for example the following code:

import "mtl" Control.Monad.State import "mtl" Control.Monad.Writer import "mtl" Control.Monad.Identity

test :: (MonadWriter [Char] m, Num t, MonadState t m) => m () test = do put 1 tell "hello"

main = do x <- return $ runIdentity $ runStateT (runWriterT test) 1 -- test :: WriterT String (StateT Int Identity) y <- return $ runIdentity $ runWriterT $ runStateT test 1 -- test :: StateT Int (WriterT String Identity) z <- runWriterT $ runStateT test 1 -- test :: StateT Int (WriterT String IO) (((), Int), String) print x print y print z

*Main> main (((),"hello"),1) (((),1),"hello") (((),1),"hello")

Until test is called in 'main' we don't know the order of monads. In fact even the base monad is not know. All we know is that it uses the State and Writer monad. In each call to 'test' in main we can determine the stacking order and the base monad yielding different results. This seems to be a more flexible way of using monad transformers but I haven't seen this in code before so is there anything wrong with this style?

-deech

_______________________________________________ Haskell-Cafe mailing list Haskell-Cafe@haskell.org http://www.haskell.org/mailman/listinfo/haskell-cafe

C. McCann

9:25 p.m.

New subject: More Flexible Monad Transformer OR Bad Coding Style

On Mon, Aug 9, 2010 at 3:42 PM, Job Vranish wrote:

...

For monads like StateT, WriterT, ReaderT, the order doesn't matter (except perhaps for some pesky performance details). However, for monad transformers like ErrorT or ListT, the order _does_ matter.

Is it really correct to say that order doesn't matter for the transformers you mention? More precise would be to say that order doesn't matter when two or more of those are stacked *consecutively*. Unless a function is completely independent of what other functions do with the state values, it can matter a great deal what order two State transformers occur in if there happens to be a ContT sandwiched between them. Furthermore, MonadState doesn't even promise that much; an arbitrary transformer that provides state operations may not "commute" generally with a StateT. Imagine, for instance, a state transformer augmented with error checking and transactions, that rolls back to a checkpoint if something 'put's an invalid state value. A polymorphic function with multiple monad typeclasses is thus effectively asserting that it does something sensible and well-defined for any set of transformers providing those classes, for any ordering of those transformers in the stack, and with any other possible transformers inside, outside, or amidst them. Combinatorics are not your friend here. Monad transformer polymorphism leads all too easily into a pit of despair. Don't go there unwisely. - C.

5447

Age (days ago)

5447

Last active (days ago)

List overview

Download

7 comments

6 participants

participants (6)

aditya siram
C. McCann
Edward Z. Yang
Gregory Crosswhite
Gábor Lehel
Job Vranish